School of Information Technology, Luoyang Normal University, Luoyang, 471934, China.
Comput Intell Neurosci. 2022 Apr 13;2022:5852891. doi: 10.1155/2022/5852891. eCollection 2022.
Lane detection, as one of the most important core functions in the autonomous driving environment, is still an open problem. In particular, pursuing high accuracy in complex scenes, such as no line and multiple lane lines, is an urgent issue to be discussed and solved. In this paper, a novel end-to-end lane detection model combining the advantages of attention mechanism and residual block is proposed to address the problem. A residual block alleviates the possible gradient problem. An attention block can help the proposed model centralize on where to focus in the process of learning feature representation, which can make the model itself more sensitive to the feature representation of lane lines through convolutional operations. Additionally, the U-shaped structure with three downsampling operations preserves the image resolution and the original lane line information in the image to the greatest extent. The U-shaped structure can directly output the prediction results to eliminate many complex or unnecessary calculation processes. The experimental results on two public lane detection datasets show that the lane detection performance of the proposed model can achieve high accuracy, and the corresponding weight sizes are only 2.25 M. Finally, to further explain the effectiveness of the proposed model, the unavoidable troubles encountered in the experiment are discussed.
车道检测作为自动驾驶环境中最重要的核心功能之一，仍然是一个悬而未决的问题。特别是在无标线和多标线等复杂场景中追求高精度，是一个亟待讨论和解决的问题。本文提出了一种新的端到端车道检测模型，结合注意力机制和残差块的优点，旨在解决该问题。残差块缓解了可能出现的梯度问题。注意力块可以帮助所提出的模型在学习特征表示的过程中集中于关注的位置，这可以通过卷积操作使模型本身对车道线的特征表示更加敏感。此外，具有三个下采样操作的 U 形结构最大限度地保留了图像的分辨率和图像中原车道线的信息。U 形结构可以直接输出预测结果，从而消除许多复杂或不必要的计算过程。在两个公共的车道检测数据集上的实验结果表明，所提出的模型的车道检测性能可以达到高精度，并且相应的权重大小仅为 2.25M。最后，为了进一步说明所提出模型的有效性，讨论了实验中不可避免的麻烦。
